Should Your AI Agent Proactively Tell a Customer About Its Own Mistake, or Wait to Be Asked

Quick Answer

Have the agent proactively disclose a mistake whenever the error could affect the customer's decision or outcome going forward, such as giving wrong pricing information they might act on. Waiting to be asked is defensible only for mistakes that are purely cosmetic or already fully corrected with no downstream impact. This is a policy question about whether to volunteer information before being asked, distinct from diagnosing whether something was a hallucination or a product bug after it has already surfaced, and from how you respond once a mistake has already gone public.

Why Silent Correction Feels Worse When Discovered

The instinct to quietly fix a mistake and say nothing is understandable: it avoids an awkward conversation and does not draw attention to the error. But when a customer discovers on their own that the agent gave them wrong information and nobody flagged it, the story they tell themselves is not "the AI made a small mistake," it is "the company knew and didn't tell me." That second story does more lasting damage to trust than the original error would have, because it reframes an isolated mistake as a pattern of concealment.

The Test: Does the Error Change What the Customer Would Do

The clearest decision rule is whether the mistake could plausibly change a customer's decision or expectation. If an agent quoted the wrong shipping estimate and the customer might have chosen a different shipping option with accurate information, that is a proactive-disclosure case. If an agent used slightly awkward phrasing that was still factually correct, or a formatting glitch that self-corrected within the same message, that does not meet the bar for proactively flagging it, since there is no decision the customer would have made differently.

What Proactive Disclosure Should Actually Look Like

A good proactive disclosure is specific and immediate: the agent states clearly what it got wrong, what the correct information is, and what if anything the customer needs to do differently as a result. It should happen as close to the moment of discovery as possible, whether that discovery comes from the same conversation or a later automated check. Vague disclosures that bury the correction in hedging language undermine the trust benefit that proactive disclosure is supposed to build.

Where Companies Get the Line Wrong

The most common mistake is over-applying this policy to genuinely trivial issues, training customers to expect frequent self-flagged corrections for things that do not matter, which erodes confidence in the agent's baseline reliability rather than building trust. The second common mistake is under-applying it to material errors out of a desire to avoid the conversation, which is the exact silent-correction pattern that backfires when discovered. The test above, whether the error changes the customer's decision, is the discipline that keeps both failure modes in check.

Building This Into Your Incident Process

Proactive disclosure should not be a case-by-case judgment call made in the moment. Build it into your standing incident postmortem and customer communication process as an explicit checklist item: for every confirmed mistake above a defined severity threshold, proactive disclosure to affected customers is the default action, not an optional extra step someone has to remember to suggest.

FAQ

Should the agent disclose the mistake itself, mid-conversation, or should a human follow up separately? For errors caught in the same conversation, the agent disclosing immediately is usually best since it is the fastest correction. For errors discovered later through auditing, a human-initiated follow-up often lands better since it signals the company is taking it seriously rather than an automated system quietly patching itself.

Does proactive disclosure create legal exposure by admitting fault? It can in some contexts, which is why the disclosure language should be reviewed alongside your broader legal risk framework rather than left entirely to the agent's own phrasing, particularly for anything touching financial or contractual commitments.

How do we know if we are over-disclosing? Track customer sentiment on disclosure messages specifically. If customers start responding with confusion or annoyance to routine disclosures, that is a signal the severity threshold is set too low.